系统级别perf top, dstat -tam, vtune 都能很好分析beam 瓶颈,本文主要erlang 级别排查:
1、反编译
确认线上运行代码是否正确,reltools没掌握好,升级偶尔出现问题
decompile(Mod) ->
{ok,{_,[{abstract_code,{_,AC}}]}} = beam_lib:chunks(code:which(Mod), [abstract_code]),
io:format("~s~n", [erl_prettypr:format(erl_syntax:form_list(AC))]).
2、进程栈
类似于jstack,发现大量进程挂起,进程数过高,运行慢,hang住等问题用到
pstack(Reg) when is_atom(Reg) ->
case whereis(Reg) of
undefined -> undefined;
Pid -> pstack(Pid)
end;
pstack(Pid) ->
io:format("~s~n", [element(2, process_info(Pid, backtrace))]).
3、etop
分析内存、cpu占用进程,即使数十w进程node 也能正常使用
%进程CPU占用排名
etop() ->
spawn(fun() -> etop:start([{output, text}, {interval, 10}, {lines, 20}, {sort, reductions}]) end).
%进程Mem占用排名
etop_mem() ->
spawn(fun() -> etop:start([{output, text}, {interval, 10}, {lines, 20}, {sort, memory}]) end).
%停止etop
etop_stop() ->
etop:stop().
4、gc all
进程内存过高时,来一发,看看是内存泄露还是gc不过来
% 对所有process做gc
gc_all() ->
[erlang:garbage_collect(Pid) || Pid <- processes()].
5、fprof
% 对MFA 执行分析,会严重减缓运行,建议只对小量业务执行
% 结果:
% fprof 结果比较详细,能够输出热点调用路径
fprof(M, F, A) ->
fprof:start(),
fprof:apply(M, F, A),
fprof:profile(),
fprof:analyse(),
fprof:stop().
6、eprof
% 对整个节点内所有进程执行eprof, eprof 对线上业务有一定影响,慎用!
% 建议TimeoutSec<10s,且进程数< 1000,否则可能导致节点crash
% 结果:
% 输出每个方法实际执行时间(不会累计方法内其他mod调用执行时间)
% 只能得到mod - Fun 执行次数 执行耗时
eprof_all(TimeoutSec) ->
eprof(processes() -- [whereis(eprof)], TimeoutSec).
eprof(Pids, TimeoutSec) ->
eprof:start(),
eprof:start_profiling(Pids),
timer:sleep(TimeoutSec),
eprof:stop_profiling(),
eprof:analyze(total),
eprof:stop().
7、scheduler usage
% 统计下1s每个调度器CPU的实际利用率(因为有spin wait、调度工作, 可能usage 比top显示低很多)
scheduler_usage() ->
scheduler_usage(1000).
scheduler_usage(RunMs) ->
erlang:system_flag(scheduler_wall_time, true),
Ts0 = lists:sort(erlang:statistics(scheduler_wall_time)),
timer:sleep(RunMs),
Ts1 = lists:sort(erlang:statistics(scheduler_wall_time)),
erlang:system_flag(scheduler_wall_time, false),
Cores = lists:map(fun({{I, A0, T0}, {I, A1, T1}}) ->
{I, (A1 - A0) / (T1 - T0)} end, lists:zip(Ts0, Ts1)),
{A, T} = lists:foldl(fun({{_, A0, T0}, {_, A1, T1}}, {Ai,Ti}) ->
{Ai + (A1 - A0), Ti + (T1 - T0)} end, {0, 0}, lists:zip(Ts0, Ts1)),
Total = A/T,
io:format("~p~n", [[{total, Total} | Cores]]).
8、 进程调度
% 统计下1s内调度进程数量(含义:第一个数字执行进程数量,第二个数字迁移进程数量)
scheduler_stat() ->
scheduler_stat(1000).
scheduler_stat(RunMs) ->
erlang:system_flag(scheduling_statistics, enable),
Ts0 = erlang:system_info(total_scheduling_statistics),
timer:sleep(RunMs),
Ts1 = erlang:system_info(total_scheduling_statistics),
erlang:system_flag(scheduling_statistics, disable),
lists:map(fun({{Key, In0, Out0}, {Key, In1, Out1}}) ->
{Key, In1 - In0, Out1 - Out0} end, lists:zip(Ts0, Ts1)).
9、trace 日志
会把mod 每次调用详细MFA log 下来,args 太大就不好看了
复制代码
%trace Mod 所有方法的调用
trace(Mod) ->
dbg:tracer(),
dbg:tpl(Mod, '_', []),
dbg:p(all, c).
%trace Node上指定 Mod 所有方法的调用, 结果将输出到本地shell
trace(Node, Mod) ->
dbg:tracer(),
dbg:n(Node),
dbg:tpl(Mod, '_', []),
dbg:p(all, c).
%停止trace
trace_stop() ->
dbg:stop_clear().
10、内存高OOM 排查工具
etop 无法应对10w+ 进程节点, 下面代码就没问题了;找到可疑proc后通过pstack、message_queu_len 排查原因
proc_mem_all(SizeLimitKb) ->
Procs = [{undefined, Pid} || Pid<- erlang:processes()],
proc_mem(Procs, SizeLimitKb).
proc_mem(SizeLimitKb) ->
Procs = [{Name, Pid} || {_, Name, Pid, _} <- release_handler_1:get_supervised_procs(),
is_process_alive(Pid)],
proc_mem(Procs, SizeLimitKb).
proc_mem(Procs, SizeLimitKb) ->
SizeLimit = SizeLimitKb * 1024,
{R, Total} = lists:foldl(fun({Name, Pid}, {Acc, TotalSize}) ->
case erlang:process_info(Pid, total_heap_size) of
{_, Size0} ->
Size = Size0*8,
case Size > SizeLimit of
true -> {[{Name, Pid, Size} | Acc], TotalSize+Size};
false -> {Acc, TotalSize}
end;
_ -> {Acc, TotalSize}
end
end, {[], 0}, Procs),
R1 = lists:keysort(3, R),
{Total, lists:reverse(R1)}.